The degradation and oxidation of fat and protein during meat storage and processing are critical factors for meat quality control and safety assessment, with pH, TVBN, and TBARS serving as key indicators of these processes. Conventional spectroscopic detection methods necessitate the training of multiple independent models to predict different attributes, a process that is both complicated and protracted. To evaluate the feasibility of non-destructive simultaneous detection of pork quality, this study developed a Multi-Task Residual 1-Dimensional Convolutional Neural Network (MTR-1D-CNN) model based on visible and near-infrared (Vis-NIR) spectroscopy for non-destructive simultaneous detection of pork samples. The MTR-1D-CNN model was compared with single-task deep learning models, including Long Short-Term Memory (LSTM) networks and 1-Dimensional Convolutional Neural Networks (1D-CNN). The predictive performance of MTR-1D-CNN for the three indicators revealed that the Rp values for each quality indicator were greater than 0.92, the R 2 p values exceeded 0.80, and the RPD values surpassed 2.5, highlighting the method's accuracy and reliability. Overall, this study not only enabled the non-destructive and simultaneous detection of pork quality indicators but also highlighted the advantages and potential applications of multi-task deep learning in food quality monitoring, providing new insights for efficient detection and quality control in the food industry. • Novel MTR-1D-CNN model designed for pork quality assessment; • Simultaneous non-destructive prediction of pH, TBARS, and TVB-N; • Multi-task learning outperforms single-task models in accuracy.
Ning et al. (Sun,) studied this question.